Microsoft Builds Its Own AI Foundation
On 2 June 2026, at its annual Build developer conference, Microsoft announced a family of seven new models developed entirely in-house by Microsoft AI. The two most significant are MAI-Thinking-1, Microsoft's first in-house reasoning model, and MAI-Code-1-Flash, a coding-specific model purpose-built for GitHub Copilot. Both were built without distillation from any third-party model — including OpenAI, with which Microsoft has a long-standing multi-billion-dollar partnership. That independence is not incidental; it is the point.
What MAI-Thinking-1 Is
MAI-Thinking-1 is a sparse Mixture of Experts model with 35 billion active parameters and approximately one trillion total parameters. It uses a 256,000-token context window and was trained from scratch on commercially licensed data, with no knowledge distilled from OpenAI, Anthropic, or any other third-party model family.
The benchmark results place it firmly in the frontier tier. On AIME 2025, the advanced mathematics competition used widely as a reasoning benchmark, MAI-Thinking-1 scores 97.0 per cent. On the harder AIME 2026 version it scores 94.5 per cent. On GPQA Diamond, the graduate-level scientific question benchmark, it scores 84.2 per cent. On SWE-bench Verified — the software engineering agent benchmark — it scores 73.5 per cent. On LiveCodeBench v6, a competitive programming benchmark, it scores 87.7 per cent. On SWE-bench Pro it scores 52.8 per cent.
Microsoft additionally reports that MAI-Thinking-1 is preferred over Claude Sonnet in blind human side-by-side evaluations — notable given that Claude Sonnet is Anthropic's mid-tier production model and one of the most widely deployed AI models in enterprise software. MAI-Thinking-1 is currently in private preview on Microsoft Foundry and available to enterprise customers upon request.
Why Trained Without OpenAI Data?
Microsoft's decision to train MAI-Base-1 — the foundational model beneath MAI-Thinking-1 — without any third-party distillation is a strategic one, not purely technical. A model trained on commercially licensed data with no external distillation carries no legal or contractual ambiguity about what a downstream customer can do with outputs at scale. For regulated industries, government contracts, and enterprise deployments where data provenance matters, that clean training lineage is a commercial differentiator.
MAI-Code-1-Flash: A New Default Inside Copilot
MAI-Code-1-Flash is a different architecture for a different purpose. It is a sparse MoE model with 5 billion active parameters and 137 billion total parameters, with the same 256,000-token context window as MAI-Thinking-1. Unlike MAI-Thinking-1, it began rolling out immediately on 2 June 2026, available across GitHub Copilot Free, Pro, Pro+, and Max plans from day one.
The model was trained directly on production Copilot harnesses and tool-use data — not on general code corpora — which means it understands how to operate within the Copilot environment's tool-calling patterns, editor integration, and multi-step agent workflows.
On SWE-Bench Pro, MAI-Code-1-Flash scores 51.2 per cent versus Claude Haiku 4.5's 35.2 per cent — a 16-percentage-point lead. On the IF Bench instruction-following benchmark, the margin is 28.9 points. The model also uses roughly 60 per cent fewer tokens than Claude Haiku 4.5 on complex coding tasks, which is directly relevant given GitHub's move to metered AI Credits billing in June 2026.
The Strategic Shift Underneath the Benchmarks
MAI-Thinking-1 and MAI-Code-1-Flash are the visible layer of a larger restructuring. Microsoft's partnership with OpenAI has been commercially transformative for both companies, but it has also created a structural dependency: every API call that Microsoft routes through its own products to OpenAI models is revenue that flows to a separate company. Building in-house models is how Microsoft begins to close that gap, the same way Google built its own TPU infrastructure rather than remaining a compute customer indefinitely.
The seven-model family announced at Build 2026 spans text, image, voice, and speech — suggesting Microsoft is building a complete in-house model stack, not just one flagship. MAI-Code-1-Flash's immediate integration into Copilot, rather than a staged preview, signals that this is product deployment, not a research announcement.
What This Means for Indian Development Teams
For Indian teams using GitHub Copilot under the AI Credits metered billing model that began in June 2026, MAI-Code-1-Flash is directly relevant to cost management. A model that uses 60 per cent fewer tokens than Haiku on equivalent tasks translates into meaningfully lower credit consumption for agentic coding sessions. Teams routing Copilot tasks through the model selector — rather than defaulting to frontier models for every request — now have a Microsoft-built option that outperforms Haiku at a more efficient token rate.
For enterprise teams evaluating Microsoft Foundry as an AI infrastructure layer, MAI-Thinking-1 arriving in private preview matters. Its commercially licensed, distillation-free training lineage addresses one of the legal concerns that has slowed some regulated Indian organisations from adopting frontier AI models wholesale. A clean data provenance story simplifies the legal review that precedes enterprise AI deployments in banking, insurance, and healthcare.
The broader implication is that AI infrastructure is no longer a market with two or three providers setting the terms. When Microsoft trains and ships its own reasoning model, it changes the competitive dynamics — and the negotiating leverage — for every enterprise that buys AI services from the major platforms.
The Bottom Line
Microsoft launched MAI-Thinking-1 and MAI-Code-1-Flash at Build 2026 on 2 June, the first in-house reasoning and coding models it has shipped without any OpenAI data. MAI-Thinking-1 scores 97 per cent on AIME 2025 and is in private preview on Microsoft Foundry. MAI-Code-1-Flash outperforms Claude Haiku 4.5 by 16 points on SWE-Bench Pro, uses 60 per cent fewer tokens on complex tasks, and is live in GitHub Copilot today. For Indian teams managing Copilot costs under metered billing, it is the most practical model upgrade available right now.
Frequently Asked Questions
What are Microsoft's MAI-Thinking-1 and MAI-Code-1-Flash models?+
MAI-Thinking-1 is Microsoft's first in-house reasoning model, a sparse Mixture of Experts model with 35 billion active parameters and a 256,000-token context window, announced at Build 2026 on 2 June 2026. It scores 97.0 per cent on AIME 2025 and 73.5 per cent on SWE-bench Verified. MAI-Code-1-Flash is a companion coding model with 5 billion active parameters, available in GitHub Copilot from 2 June 2026 across all plan tiers — Free, Pro, Pro+, and Max.
How does MAI-Thinking-1 compare to other frontier AI models on benchmarks?+
On AIME 2025, MAI-Thinking-1 scores 97.0 per cent; on AIME 2026 it scores 94.5 per cent. On GPQA Diamond it scores 84.2 per cent, on LiveCodeBench v6 it scores 87.7 per cent, and on SWE-bench Verified it scores 73.5 per cent. Microsoft additionally reports it is preferred over Claude Sonnet in blind human side-by-side evaluations. The model is in private preview on Microsoft Foundry.
How does MAI-Code-1-Flash compare to Claude Haiku 4.5 for coding tasks?+
On SWE-Bench Pro, MAI-Code-1-Flash scores 51.2 per cent versus Claude Haiku 4.5's 35.2 per cent — a 16-percentage-point advantage. On the IF Bench instruction-following benchmark the margin is 28.9 points. The model also uses approximately 60 per cent fewer tokens than Haiku on complex coding tasks, which directly reduces credit consumption under GitHub Copilot's metered AI Credits billing.
Why did Microsoft train its MAI models without OpenAI data?+
Microsoft trained MAI-Thinking-1 on commercially licensed data with no knowledge distillation from any third-party model, including OpenAI. This gives the model a clean data provenance record — an important factor for regulated enterprises in banking, healthcare, and government that need clarity on training-data licensing before deploying AI at scale. It also reduces Microsoft's commercial dependency on OpenAI, beginning to close the revenue transfer embedded in their long-standing AI partnership.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.